Bass Diffusion Model and data clustering based forecasting method for fashion goods: A case study

Bass diffusion model (BDM) is known as one of the successful models when it comes to new product adoption. In this study, using real world sales data, we analyzed the performance of BDM in reflecting sales curve of fashion type products. The study consists of two main parts; first we fit BDM by using all data points to investigate whether BDM is flexible enough to represent the pattern of sales or not. Then, we used products' partial data to find out the forecasting performance of BDM in forecasting the remaining demand of product in a season. We observed that BDM with partial data can not give satisfying results. Therefore, in the second part of our study, we focus on developing a successful forecasting approach and we developed methods based on BDM based on partially observed sales data in a current season, and clustering algorithm on full data sales data sets observed in the past. We analyzed the performance of several proposed forecasting methods. The performance of BDM is highly dependent on the accuracy of parameters estimation. Three main parameter estimation methods are used in the literature. These are the ordinary least squares method, the non-linear least squares method, and the maximum likelihood method. We utilized these methods in the study and compare them. In addition, we propose a new approach, we call randomized line search method. This method gives comparable results with the other parameter estimation techniques. As a consequence of first part of our work, we observe that BDM is capable of representing the sales of fashion products regarding to remaining demand of a product in a season. But with partial data BDM does not produce accurate results, especially with limited data from early periods in a season. To handle this problem, we developed a new model which relates past products to the new products when creating forecasts for a new product. Clustering methods, regression analysis and time-series analysis are utilized in development of the new forecast model. First, past sales data of products are clustered according to their sales curves. Each cluster represents different sales pattern. New product with a partial data is compared with the past products and the cluster with the minimum dissimilarity is selected. Using the cluster information, the "closest" product information and the partial data of the new product, a forecast remaining sales in a season is created for the new product . We suggested different ways to set weights for the information gained from clusters, from the closest product and from the new product. We created forecasts with proposed model and the BDM, and compared the forecasting performance of the suggested methods. We showed that some of the suggested methods have satisfying forecasting performance.